Heads up: posts on this site are drafted by Claude and fact-checked by Codex. Both can still get things wrong — read with care and verify anything load-bearing before relying on it.
why → how

A short history of AI, from Turing to today's LLMs

Seventy years of trying to make machines think — and how a single architecture from 2017 finally cashed the check that 1950s AI wrote.

AI & ML intermediate Apr 29, 2026

Why it exists

If you only encountered AI in the last few years, the story looks like a flash: ChatGPT appeared in late 2022 and within a year everyone was talking about it. But the idea of a thinking machine is older than the transistor, and the people who chased it spent most of the intervening decades being wrong in interesting ways.

The history matters because almost every “new” idea you’ll read about — agents, reasoning, alignment, even the fear of superintelligence — was proposed, tried, abandoned, and revived at least once before. Knowing the arc helps you tell the genuinely new ideas from rebrands, and it explains why the current moment feels both inevitable and surprising at the same time.

Why it matters now

The current crop of LLMs inherited their architecture, their training tricks, and most of their vocabulary from work done long before the GPT era. The reason today’s models work isn’t a single breakthrough — it’s the accumulation of about five generations of partial successes plus, finally, enough compute and data to make the whole stack pay off.

Understanding that lineage is the difference between treating AI as magic and treating it as engineering. It also tells you where the next bend is likely: the field has a habit of solving a long-standing problem and then immediately discovering that the solution unlocks a new one.

The short answer

AI history ≈ symbolic logic → statistical learning → scaled neural nets

For seventy years, AI swung between two big ideas: write down rules that encode human reasoning (symbolic AI), or have machines learn patterns from data (statistical / connectionist AI). The second idea kept losing because the hardware and datasets weren’t there. Around 2012 they finally were, and the rest of the story is what happens when one approach starts working faster than anyone can keep up with.

How it works

A rough timeline, with the load-bearing ideas called out.

1950 — Turing’s question. Alan Turing publishes Computing Machinery and Intelligence, asks “can machines think?”, and proposes the imitation game as a way to dodge the philosophy and get to engineering. No machine could play it for decades, but the framing stuck.

1956 — Dartmouth, and the name “AI”. A summer workshop at Dartmouth coins the term artificial intelligence and sets the optimistic tone: participants believed a serious dent in the problem was a few years off. Early systems could prove geometry theorems and play checkers.

1960s–70s — Symbolic AI. The dominant bet was that intelligence = symbol manipulation. If you could express knowledge as rules and run logic over them, you’d get reasoning. This produced expert systems that worked impressively in narrow domains (MYCIN for infections, DENDRAL for chemistry) and failed completely outside them.

Meanwhile — the connectionist underdog. Frank Rosenblatt’s perceptron (1958) showed a single-layer neural network could learn simple patterns from examples. Minsky and Papert’s 1969 book Perceptrons proved it couldn’t learn things as basic as XOR, and funding for neural nets evaporated. The first AI winter followed.

1986 — Backpropagation, take two. Rumelhart, Hinton, and Williams popularize backpropagation, which lets multi-layer networks actually learn. In principle this answered Minsky’s objection. In practice the networks were tiny, slow, and beaten by simpler statistical methods. A second AI winter arrived in the early 90s when expert systems also failed to scale.

1990s–2000s — Statistical machine learning takes over. Translation, spam filtering, search ranking, speech recognition — all the practical “AI” wins came from classical machine learning: support vector machines, random forests, hidden Markov models. Quietly useful, not glamorous.

2012 — AlexNet and the GPU moment. A neural network called AlexNet wins the ImageNet image-recognition contest by a huge margin, using GPUs to train a model nobody could have trained on CPUs. Deep learning — neural nets with many layers — suddenly works. Within three years it dominates vision, speech, and translation.

2014–2016 — Sequence models. RNNs and especially LSTMs become the standard for text and speech. Useful, but they process tokens one at a time, which is slow to train and bad at long-range dependencies.

2017 — Attention is all you need. A Google paper introduces the transformer. Instead of stepping through a sequence, every token attends to every other token in parallel. This is the unlock: it trains far better on GPUs and captures long-range structure that LSTMs missed. Almost every modern AI model — text, image, audio, video — is now a transformer or a close cousin.

2018–2020 — GPT and the scaling laws. OpenAI’s GPT-1, GPT-2, GPT-3 demonstrate something quietly radical: if you make the model bigger, feed it more text, and train longer, capabilities improve predictably. GPT-3 (175 billion parameters, 2020) could write essays, code, and pastiches with no task-specific training. The scaling laws turned AI progress from research lottery into capital expenditure.

2022 — ChatGPT and RLHF. GPT-3 was powerful but raw. ChatGPT bolted on RLHF: fine-tune the base model to follow instructions, then use human ratings to shape its behavior into a helpful, mostly-honest assistant. This wasn’t a capability leap so much as a product leap — and it broke containment. A hundred million users in two months.

2023–2025 — Multimodality, tools, and agents. Models learn to see images (GPT-4V, Claude 3), generate them (DALL·E, Midjourney, Stable Diffusion), and use external tools by calling them as functions. This is when the agent pattern crystallizes: an LLM in a loop, deciding what to do next and calling tools to do it. See agent.

2024–2026 — Reasoning and long horizons. OpenAI’s o-series and DeepSeek-R1 introduce models that think before answering — generating long internal chains of reasoning and being trained, via reinforcement learning, to do this well. Context windows reach a million tokens; agents start completing work that takes hours instead of seconds. Open-weights models catch up with frontier closed ones on most benchmarks. The current frontier is no longer “can the model answer?” but “can it stay coherent across a day of work?”

Going deeper